Research map / Architectures and efficiency / Pruning and efficient tuning
OrthoFuse: Training-free Riemannian Fusion of Orthogonal Style-Concept Adapters for Diffusion Models
Ali Aliev, Kamil Garifullin, Nikolay Yudin, Vera Soboleva, Alexander Molozhavenko, Ivan Oseledets, Aibek Alanov, Maxim Rakhuba
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| merge style and concept adapters with OrthoFuse instead of K-LoRA or ZipLoRA | SDXL orthogonal adapters · concept-style fusion | CLIP style similarity: 0.49 -> 0.61 | Improved |
| merge with OrthoFuse instead of K-LoRA or ZipLoRA | SDXL · concept-style fusion | concept CLIP / DINO similarity: CLIP 0.76 -> 0.68; DINO 0.64 -> 0.51 | Hurt |
| merge with OrthoFuse instead of K-LoRA r = 64 | SDXL · concept-style fusion | geometric mean (style, CLIP): 0.61 -> 0.64 | Improved |
| fuse training-free with OrthoFuse instead of joint orthogonal training | SDXL · concept-style fusion | CLIP style similarity: 0.48 -> 0.61 | Improved |
| merge training-free instead of joint training | SDXL · concept-style fusion | merging time: 1.5 hours -> < 1 sec | Improved |
| increase LoRA rank of baselines from 8 to 64 | ZipLoRA · concept-style fusion | DINO: 0.55 -> 0.64 | Improved |
| increase LoRA rank of baselines from 8 to 64 | K-LoRA · concept-style fusion | CLIP style similarity: 0.46 -> 0.49 | Improved |
| add spectra restoration (eigenvalue rotation) after geodesic interpolation (ablation) | OrthoFuse · SDXL concept-style fusion | style expressiveness (eigenvalue spectrum) | Improved |
| set fusion parameter t to 0.6 (ablation) | OrthoFuse · SDXL concept-style fusion | style-concept balance: t = 0.6 | Improved |
| set phase multiplier midpoint phi0 = 2 (ablation) | OrthoFuse · SDXL concept-style fusion | fusion quality: phi0 = 2 | Improved |
| compare against K-LoRA in human preference on concept preservation | OrthoFuse · user study (65 participants, 1460 comparisons) | concept preservation preference: 48% vs 52% | No clear effect |
Closest papers by meaning
- Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes (ICML 2024 · similarity 0.87)
- Orthogonal Adaptation for Modular Customization of Diffusion Models (CVPR 2024 · similarity 0.85)
- Efficient Diffusion Models for Symmetric Manifolds (ICML 2025 · similarity 0.85)
- StelLA: Subspace Learning in Low-rank Adaptation using Stiefel Manifold (NeurIPS 2025 · similarity 0.85)
- Manifold Diffusion Fields (ICLR 2024 · similarity 0.85)
- Generalizing the Geometry of Model Merging Through Fréchet Averages (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
